Re: [PATCH] perf: enum overflow in uapi/linux/perf_event.h

Re: [PATCH] perf: enum overflow in uapi/linux/perf_event.h

[Christophe LEROY]

Le 07/09/2018 à 15:58, Peter Zijlstra a écrit :
> On Fri, Sep 07, 2018 at 01:50:18PM +0000, Christophe Leroy wrote:
>>
>>
>> On 09/07/2018 01:42 PM, Peter Zijlstra wrote:
>>> On Fri, Sep 07, 2018 at 01:27:19PM +0000, Christophe Leroy wrote:
>>>> On PPC32, enums are 32 bits, so __PERF_SAMPLE_CALLCHAIN_EARLY is
>>>> out of scope. The following sparse warning is encountered:
>>>>
>>>>     CHECK   arch/powerpc/kernel/process.c
>>>> ./include/uapi/linux/perf_event.h:147:56: warning: cast truncates bits from constant value (8000000000000000 becomes 0)
>>>
>>> Urgh... what compiler is that? I've not seen anything like that from the
>>> build bots.
>>>
>>
>> [root@pc16082vm linux-powerpc]# sparse --version
>> 0.5.2
>>
>> [root@pc16082vm linux-powerpc]# ppc-linux-gcc --version
>> ppc-linux-gcc (GCC) 5.4.0
> 
> Ah, that's a sparse warning. But does your GCC agree? The thing is,
> sparse uses the C enum spec, but I suspect GCC uses the C++ enum spec
> and it all works fine.
> 

Ah yes, it seems that GCC is happy. So sparse should be fixed instead ?

Anyway, is it really correct to put this constant inside that enum, 
after PERF_SAMPLE_MAX  ?

Christophe

Re: [PATCH] perf: enum overflow in uapi/linux/perf_event.h

[Peter Zijlstra]

On Fri, Sep 07, 2018 at 04:15:33PM +0200, Christophe LEROY wrote:

> Ah yes, it seems that GCC is happy. So sparse should be fixed instead ?

Ideally, yes.

> Anyway, is it really correct to put this constant inside that enum, after
> PERF_SAMPLE_MAX  ?

It is a bit of a hack, agreed. What we do is use the top bit of that
word (u64) for some internal state. By placing it there (after MAX) we
ensure it is not available for userspace (trying to set it will return
in -EINVAL) and by keeping it in the enum we know that bit is
unavailable for future use.

I have a patch queued that puts a little comment on that:

  https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/commit/?h=perf/urgent&id=34cad593c9ea350a1811ab718e64b36e5cde870c

(url is not stable, as I regenerate that git tree from quilt every so
often, but it should probably last the day).

Re: [PATCH] perf: enum overflow in uapi/linux/perf_event.h

[Luc Van Oostenryck]

On Fri, Sep 07, 2018 at 04:15:33PM +0200, Christophe LEROY wrote:
> Le 07/09/2018 à 15:58, Peter Zijlstra a écrit :
> > On Fri, Sep 07, 2018 at 01:50:18PM +0000, Christophe Leroy wrote:
> > > 
> > > 
> > > On 09/07/2018 01:42 PM, Peter Zijlstra wrote:
> > > > On Fri, Sep 07, 2018 at 01:27:19PM +0000, Christophe Leroy wrote:
> > > > > On PPC32, enums are 32 bits, so __PERF_SAMPLE_CALLCHAIN_EARLY is
> > > > > out of scope. The following sparse warning is encountered:
> > > > > 
> > > > >     CHECK   arch/powerpc/kernel/process.c
> > > > > ./include/uapi/linux/perf_event.h:147:56: warning: cast truncates bits from constant value (8000000000000000 becomes 0)
> > > > 
> > > > Urgh... what compiler is that? I've not seen anything like that from the
> > > > build bots.
> > > > 
> > > 
> > > [root@pc16082vm linux-powerpc]# sparse --version
> > > 0.5.2
> > > 
> > > [root@pc16082vm linux-powerpc]# ppc-linux-gcc --version
> > > ppc-linux-gcc (GCC) 5.4.0
> > 
> > Ah, that's a sparse warning. But does your GCC agree? The thing is,
> > sparse uses the C enum spec, but I suspect GCC uses the C++ enum spec
> > and it all works fine.

Sparse is a bit weird about the exact underlying type used for enums.

> Ah yes, it seems that GCC is happy. So sparse should be fixed instead ?

I'll investigate (I suppose the same is given on x86-32).

-- Luc

Re: [PATCH] perf: enum overflow in uapi/linux/perf_event.h

[Luc Van Oostenryck]

On Fri, Sep 07, 2018 at 08:43:59PM +0200, Luc Van Oostenryck wrote:
> On Fri, Sep 07, 2018 at 04:15:33PM +0200, Christophe LEROY wrote:
> > Le 07/09/2018 à 15:58, Peter Zijlstra a écrit :
> > > On Fri, Sep 07, 2018 at 01:50:18PM +0000, Christophe Leroy wrote:
> > > > 
> > > > 
> > > > On 09/07/2018 01:42 PM, Peter Zijlstra wrote:
> > > > > On Fri, Sep 07, 2018 at 01:27:19PM +0000, Christophe Leroy wrote:
> > > > > > On PPC32, enums are 32 bits, so __PERF_SAMPLE_CALLCHAIN_EARLY is
> > > > > > out of scope. The following sparse warning is encountered:
> > > > > > 
> > > > > >     CHECK   arch/powerpc/kernel/process.c
> > > > > > ./include/uapi/linux/perf_event.h:147:56: warning: cast truncates bits from constant value (8000000000000000 becomes 0)
> > > > > 
> > > > > Urgh... what compiler is that? I've not seen anything like that from the
> > > > > build bots.
> > > > > 
> > > > 
> > > > [root@pc16082vm linux-powerpc]# sparse --version
> > > > 0.5.2
> > > > 
> > > > [root@pc16082vm linux-powerpc]# ppc-linux-gcc --version
> > > > ppc-linux-gcc (GCC) 5.4.0
> > > 
> > > Ah, that's a sparse warning. But does your GCC agree? The thing is,
> > > sparse uses the C enum spec, but I suspect GCC uses the C++ enum spec
> > > and it all works fine.
> 
> Sparse is a bit weird about the exact underlying type used for enums.
> 
> > Ah yes, it seems that GCC is happy. So sparse should be fixed instead ?
> 
> I'll investigate (I suppose the same is given on x86-32).


It's definitively a bug in sparse. A relatively nasty one and which
open a can of worms. Fortunately, I had already looked at these
problems in May, I just didn't had the time to push the patches.

-- Luc